A Multiagent System using Associate Rule Mining (ARM), a collaborative filtering
approach
Prakash Ranganathan, Juan Li, Kendall Nygard
Department of Computer Science
North Dakota State University, Fargo, ND 58102
email:prakashranganathan@mail.und.edu
Abstract___Agent Oriented Programming (AOP) is a recent
promising software paradigm that brings concepts from the
theories of artificial intelligence into the mainstream realm
of distributed systems, and yet it is rather difficult to find a
successful application of agent oriented system (specifically)
when large-scale systems are considered. When adopting an
agent-oriented approach to solve a problem, there are a
number of domain independent issues that must always be
solved, such as how to model agent behavior to predict
future action and how to allow agents to communicate
rather than expecting developers to develop this core
infrastructure themselves. In our paper, we address several
problems that exist in a socialized e-learning environment
and provide solutions to these problems through smart and
collaborative agent behavior modeling which learn and
adapt themselves through prior experiences, thereby
assisting in successful implementation of this large scale e-
learning system. In this paper, the author (s) proposes an
implementation of a complete distributed e-learning system
based on Collaborative filtering (CF) method. The system
has intelligent collaborative filtering based tutoring system
(ICFTS) capabilities that allow contents, presentation and
navigation to be adapted according to the learner’s
requirements. In order to achieve that development, two
concepts were put together: multi-agent systems and data
mining techniques (specifically, the ARM algorithm). All the
implementation code is developed using MATLAB GUI
environment.
To our best knowledge, very few literatures discusses a
portion of e-learning environment using adaptive software
agents, but none of the current literatures addresses a
complete implementation of their learning system in detail.
The goal of the paper is to implement one such multi-agent
based e-learning system which learns from its prior user
experiences on top of an agent-oriented middleware that
provides the domain-independent infrastructure, allowing
the developers to focus on building the key logic behind it. In
this system,
the agents follow an adaptive cognitive learning approach,
where the agent learns through user
behaviors via a collaborative filtering technique, or
experiencing and then processing and remembering the
information in an e-learning environment. The paper will
utilize agent (a piece of code) based environment in our e-
learning system using ARM [1][2]. The paper follows a
learning approach based cognitive domain of Bloom's
Taxonomy such as Analyze, Evaluate, Create, Apply,
understand and remember.
c
978-1-4244-6349-7/10/$26.00 2010 IEEE
Keywords: Collaborative filtering, intelligent agents
I. INTRODUCTION
The concept and practice of distance education has
evolved tremendously at numerous educational
institutions and are growing exponentially in recent years
around the world and yet there is not much research done
to employ adaptive learning agents or algorithms which
can predict the behavior of student users to make their
learning process easier in an e-learning environment.
Current e-learning system in literatures does not cite any
use of rigorous adaptive software agents in detail to study
the behavior of student users in promoting their
knowledge base or assist in their e-learning environment.
Effective e-learning systems should include advanced
functions with adaptive learning methods and their
interface should hide their complexity to learners,
providing an easy interaction grasping the student’s
interest. Despite of this, we often find a mere electronic
transposition of traditional material, provided through
rigid interaction schemes and awkward interfaces.
In our paper we follow a Market basket approach, but
here the customers are students in e-learning environment
and the products they shop are the query they visit in an
online interface, So, we call our approach as Query
Basket Analysis, as seen in Fig.1 below. Frequent item set
mining leads to the discovery of associations and
correlations among items in large transactional or
relational data sets. With massive amounts of data
continuously being collected and stored, many industries
are becoming interested in mining such patterns from
their databases. The discovery of interesting correlation
relationships among huge amounts of business (query)
transaction records can help in many e-learning (business)
decision making processes, such as course design, cross
marketing, and student shopping-behavior-analysis.
In our paper, we present an agent oriented e-learning
framework employing adaptive software agents, which
learns from user behavior (via a participation of rating
scheme) who share similar experience in their learning
using collaborative filtering technique [5][6]. In our
method, the student users rates the difficulty or
V7-574
completion level of the questions in a course using a
rating scale ranging from 0 to 5 with 0 (1 or 2 or 3 or 4)
being higher difficulty level in understanding the problem
signifying the student need more attention or help in
solving the problem further and 5 being a rating where
student user can easily or moderately understand the
problem stated.
Which questions are
frequently visited by
my students?
Q1 Q2
Q3 Qn
ĞͲŶĂůLJƐƚ
Fig.1 Query (Market) Basket Analysis 
We focus on a 5 star rating scheme in our problem to
filter out students who are not having trouble in solving
the problem. These students experience are shared to
determine and predict whether the next student will
complete the query in the course interface. Then, a
recommendation agent is deployed to the student interface
for student users who are unable to complete the question,
by providing more help to the student associated with the
particular problem that he or she having difficulty in
understanding the query or the problem. The
recommendation agent deploys various e-assist tools on
the problem to the student agent, such as an e-help
window on the particular problem, an instructor – student
chat interface to further assist the student [7] seen in Fig
2. Thus provide a high degree of adaptive intelligent
learning mechanism according to their logged history of
attended student user queries.
Fig.2 Use-case diagram of a Student User
[Volume 7] 2010 2nd International Conference on Computer Engineering and Technology
Our collaborative filtering algorithm uses a collection of
student user profiles to identify “interesting information"
for these users. A particular student user gets a
recommendation based on the user profiles of other,
similar users. User profiles are commonly obtained by
explicitly asking users to rate the items. Collaborative
filtering has often been formulated as a
Self-contained problem, apart from the classic
information retrieval problem (i.e. ad hoc text retrieval).
The rest of the paper is divided into following sections:
GUI Introduction and Learning Interface, Agent behavior
model, Query Association Rule and Conclusion
Fig.3 A generic e-learning system architecture
II. E-LEARNING INTERFACE
Our MATLAB based GUI interface has question set on
the right side relating to a course and five agents at the
entry login phase such as Instructor Agent, Behavior or
Recommendation Agent, Prospective Student Agent,
Directory agent, Grader agent and Service agent. We will
primarily focus on Recommendation agent employing
Collaborative filtering method to study the behavior of the
users in the system.
Fig.4 A model of MATLAB/JADE Interface representing multi-agent e-
learning System
V7-575

The Foundation for Intelligent Physical Agents (FIPA) is
an international organization that is dedicated to
promoting the industry of intelligent agents by openly
developing specifications to support interoperability
amongst agents and agent-based systems. JADE is a FIPA
compliant agent platform and development framework.
Agent platforms are responsible for dealing with agent
services such as messaging (transport, encoding and
parsing), scheduling, agent lifecycle management and
other common resources. JADE agents are executed in a
container where they share these services with other
agents present in the container. Our main interest in the
JADE platform is to build a adaptive learning agent and
implement an multi-agent oriented framework[3][4].

III. III. AGENT BEHAVIOR MODEL USING
COLLABORATIVE FILTERING (CF)
Collaborative-filtering algorithms aim to identify users
that have relevant interests and preferences by calculating
similarities and dissimilarities between user profiles
(Herlocker et al., 2004) [9]. The idea behind this method
is that, it may be of benefit to one’s search for information
to consult the behavior of other users who share the same
or relevant interests and whose opinion can be trusted. So,
our collaborative filtering technique assumes the
following: 1. Student Users who were behaved (rated)
similar in the past are likely to be behaving similar in the
future. 2. Use similar users’ behaviors to make
recommendations [4].3. All student user must have a
fixed timeframe (say 15 minutes) in responding to each
question on the agent interface.4. All e-assist tools are
deployed ahead (say 3 minutes) before the closing of each
query session.5. We assume a predefined behavior
through a behavior rating Table containing 4 users and 5
sample queries in a course.
The purpose of our e-learning system is to learn the
behavior of each student user through their responses to a
set number of questions and adapt to offer a personalized
solution through agent oriented programming
automatically, if the student had problem in understanding
the question. We decide to use association rule mining
(ARM), since it discovers associations between sets of
virtual agent communities that are shared across many
student users [2] [9]. By virtual community we mean "a
group of agents who share characteristics and interact in
essence or effect only". In other words, people in a virtual
community influence each other as though they interacted
but they do not interact. To carry out this agent based
adaptive learning task, we developed the e-learning
interface using MATLAB GUI environment. Our e-
learning system assumes a pre-defined set of questions
(five questions, in our example) pertaining to a problem in
V7-576 2010 2nd International Conference on Computer Engineering and Technology
a course. For our simplicity, we have adapted a Data
Structure course in C++ programming environment. Our
paper describes an ARM implementation in MATLAB
GUI/JADE using CF.



Fig.5 Behavior Prediction using Collaborative filtering
A. Association Rule Mining (ARM)
Association rule mining consists of first finding frequent
item sets (set of items such as A and B, satisfying a
minimum support threshold, or percentage of the task-
relevant tuples) from which strong association rules in the
form of A=>B are generated. These rules also satisfy a
minimum confidence threshold (a pre-specified
probability of satisfying B under the condition that A is
satisfied). Associations can then be further analyzed to
uncover correlation rules, which convey statistical
correlations between item sets A and B. Recommendation
agents need to employ efficient prediction algorithms so
as to provide accurate recommendations to users. If a
prediction is defined as a value that expresses the
predicted likelihood that a user will ‘‘like’’ an item, then a
recommendation is defined as the list of n items with
respect to the top-n predictions from the set of items
available. Improved prediction algorithms indicate better
recommendations. This explains the essentiality of
exploring and understanding the broad characteristics and
potentials of prediction algorithms and the reason why
this work concentrates on this research direction. In our
paper, we use the ARM based CF technique to predict the
student user behavior and make recommendation to
support a student or not in their learning process.
IV. IV. QUERY ASSOCIATION RULE
(QAR), SUPPORT AND CONFIDENCE
RULE
We define some definitions for Associative Rule Mining
on Student Collaborative Filtering (SCF) data, such as the
entities U (i.e., the set of Student Users ) and q (i.e., the
set of queries or questions of multiple choice type in a
course), and for any subset of rating values, R ⊂ {0,1,2,3,
[Volume 7]
 4, 5}, a R refers to a rating relationship relating user, u, to
question q, iff u rates q with a rating from R, e.g., for
R={0,1,2,3,4,5} in RT listed below. If a user rates to a
question, then he has attended and successfully responded q1 q2 q3 q4 q5 Ui
to the question and it is assigned a 1 in the Routing Table
or else a zero as shown in the routing table. It is important
to note that a 0 in the Routing Table implies that the U1 1 0 0 0 0 1
student user either not visited this question on the GUI
interface or visited, but failed to complete successfully or U2 1 1 0 1 0
the user may have rated 1, 2, 3 or 4 rating for the query. 3
Let A, be a set of questions, then the User set of A with
respect to R, UsetR(A), is the set of users, u, such that U3 1 0 1 0 0 2
∀a∈A, u Rrates a [(UsetR({q1,q2})= {u2,u4})]. Similarly,
the User count of A w.r.t R, UcountR(A), is the count of u U4 1 0 1 0
such that ∀a∈A, u Rrates a ? 3
[(UcountR({q1,q2})=|{u2,u4}|=2)]. User ratio of A w.r.t R,
UratioR(A), is the ratio of u such that ∀a∈A, u Rrates a. mũ 4 2 1 2 0
[(UratioR{q1,q2}=UcountR{q1,q2}/4=2/4)].

TABLE 1: BEHAVIOR ROUTING TABLE

u ∈QsetR(u)
ratings for q∈ User rates 5* , System assigns a 1 in the
BRT to represent query completed
U1 r(u1,q1) User rates 0,1,2,3 or 4 - query not visited/
U2 r(u2,q1) r(u2,q2) r(u2,q4) not completed or having difficulty, the
U3 r(u3,q1) r(u3,q3) system assigns 0 for all of the above ratings
U4 r(u4,q1) r(u4,q2) r(u4,q4)
Fig. 7 (a and b): User- Query Behavior Ratings Table (BRT)

Goal: Predict the behavior of student user SU4 in visiting query

q2 (i.e., rate) in the JADE/GUI interface?

Notes on Fig 7.a: 1 means that the user has chosen this query
and completed the query successfully in the JADE/GUI interface
and a 0 means that the user has not chosen this query or not
completed the question successfully in the JADE/GUI interface

We mainly rely on two factors to make a predictive

Fig.6
Routing Table Graph
We develop a new Query Association Rule (a QAR),
where A→C, associates two disjoint sets A and C ⊂ Q.
We form new set A called the antecedent, because it is
listed first in the RT graph and a set C, the consequent, as
it listed last in RT graph.
decision using the minimum Support (Supportmini) and
Confidence rule to understand the behavior of student
learners in the JADE interface.
The User set [count] {ratio=support} of QAR, AÆC,
w.r.t R can be written as, UsetR(AÆC) [UcountR(AÆC)]
or {UratioR(AÆC)}, is that of A∪C. Note: In ARM,
UratioR(AÆC) is called “support” of AÆC,
UsupR(AÆC). The User confidence of a QAR, AÆC,
w.r.t R, UconfR(AÆC) can be thought as
UsupR(A∪C)/UsupR(A) or countR(A∪C)/UcountR(A).
ARM Data Miners typically want to find all Rstrong
Rules, A→C, i.e., those with UsupR(A→C) > Supportmini
and UconR(A→C) > Confidencemini
(i.e. QsupR(AÆC) > Supportmini QsupR(AÆC) >
Confidencemini) with A and C form a Rstrong Correlation,
A↔C, if both AÆC and CÆA are Rstrong rules. So for
a particular q, look for n such that UcountR(m,n) >
UcountR(n)* Confidencemini. It is to be noted that

[Volume 7] 2010 2nd International Conference on Computer Engineering and Technology V7-577
UcountR(n) > UcountR(m,n) at always. So for the
example at right, let’s do ARM to predict (u,q) = (u4,q2).
We look for q ≠ n ∈ QsetR(u4) such that
Ucount(q2,n)>Ucount(n) * (7/8). We find the Ucount’s
values are taken with u4 removed, so we get a user chosen
number for Confidencemini = 7/8.
Using the condition: Ucount(q2,n)>Ucount(n)
* (7/8)
n=q1, 1 > 3 * (7/8) FALSE
n=q4, 1 > 1 * (7/8) TRUE
Fig.8 User Confidence condition
So {q4}Æ{q2} is a confident rule and therefore, likely,
since q4 rates u4, q2 likely will too. As we know now that
the student user U4 will visit and complete q2, then this
user need not need any help. On the other hand, if the
value turn to be 0 i.e., had the confidence rule failed or
student not visited q2, then the recommendation agent
will be automatically called up into the GUI interface to
assist the student in wondering about query-2 and provide
personalized assistance in completing the problem,
thereby creating an adaptive learning environment using
CF approach.
V. V. RESULTS
discussed an implementation of ARM based Collaborative
filtering in e-learning environment using MATLAB
interface. A work in progress on similar implementation
to test them in a JADE interface for verification with our
preliminary confidence metric results large data set with
detailed semantic rules.
REFERENCES
[1] FIPA, "FIPA Agent Message Transport Service Specification," 2002.
[2] Charu C. Aggarwal, Zheng Sun, and Philip S. Yu, Online Algorithms
forFinding Profile Association Rules, Proceedings of the ACM CIKM
Conference, 1998, pp 86-95
[3] FIPA, "FIPA Agent Message Transport Protocol for WAP
Specification," 2002.
[4] FIPA, "FIPA Agent Message Transport Protocol for IIOP
Specification," 2002.
[5] Qualitative analysis of user-based and item-based prediction
algorithms for recommendation agents, Manos Papagelisa, Dimitris
Plexousakisa, ELSEVIER, Engineering Applications of Artificial
Intelligence 18 (2005) 781–789.
[6] CSCI783/785, Dr. William Perizzo Notes in High Performance
Computing, Computer Science, North Dakota State University
[7]. B. Sarwar, G. Karypis, J. Konstan, J. Riedl, Analysis of
recommendation algorithms for e-commerce, in: Proceedings of ACM
ECommerce,2000
[8]. H. Nwana, D. Ndumu, A perspective on software agents research,
Knowledge Engineering Review 14 (2) (1999) 1–18.
[9] Herlocker, J., Konstan, J.A., Terveen, L., Riedl, J., 2004. Evaluating
collaborative filtering recommender systems. ACM Transactions on
Information Systems (TOIS).

We have tested the query rules with smaller data set of 10

students and 5 queries (multiple runs), which is pretty
small on MATLAB GUI environment using Associative
Rule Mining (ARM), but our result show a confidence
and support metric greater than 60% for each user
participation reflecting the potential of using ARM based
environment in future e-learning system.

VI. VI. CONCLUSION

It is evident that agent based learning object repositories

are part of the future of e-Learning system, because of the
great level of potential for cost reduction and personalized
instruction[8]. The approach presented, could be
expanded easily to each course being considered as a
learning object, with a objective description of the course
given by its title, number, list of pre-requisites and so on.
We do see several challenges before learning objects can
fulfill their potential, especially large data set for
automated deployment of e-assist tools. We believe that
collaborative filtering and rules has good potential in this
kind of learning environment. We have demonstrated and

V7-578 2010 2nd International Conference on Computer Engineering and Technology [Volume 7]